Model training and content recall methods, media, apparatuses, and computing devices

Abstract

Embodiments of the present disclosure provide a model training and content recall method, medium, device and computing device. The model training method comprises: obtaining a plurality of first samples and a plurality of second samples; training a first factor decomposition machine model according to the first samples to obtain a trained first factor decomposition machine model, wherein the first samples comprise at least one user portrait of a sample user, a content identifier of a sample content and corresponding sample click information, and the sample click information is used to indicate that the sample user has clicked the sample content or has not clicked the sample content; and training a second factor decomposition machine model according to the second samples to obtain a trained second factor decomposition machine model, wherein the second samples comprise a user identifier of a sample user, at least one content label of a sample content and sample click information. Training in combination with identifier and label type features enables the trained model to have better generalization, thereby improving the accuracy of recalled content.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the field of artificial intelligence technology, and more particularly, embodiments of the present disclosure relate to a model training and content recall method, medium, device and computing device. BACKGROUND

[0002] This section is intended to provide background information to facilitate an understanding of embodiments of the disclosure set forth in the claims. The description herein does not constitute admission of prior art.

[0003] The emergence and popularity of the Internet have brought a large amount of information to users. In the information age, in order to meet the needs of users for information, it is necessary to filter out useful information for users from a large amount of information and recommend it to users.

[0004] When a recommendation system recommends content to a user, it mainly involves two stages, namely, recall and sorting. The recall stage mainly selects a batch of content to be recommended from a large amount of content, and the sorting stage sorts the batch of content to be recommended, thereby determining the content to be recommended to the user. Since the content to be recommended is determined from the batch of content to be recommended in the recall stage, the accuracy of recall is crucial to the final recommendation.

[0005] The current recall scheme is mainly completed based on a recall model. The recall model is usually trained by using user identifiers and content identifiers. Since different users and different content correspond to different user identifiers and content identifiers, the recall model trained by using the user identifiers and the content identifiers has poor generalization, and the accuracy of recall is low. SUMMARY

[0006] The present disclosure provides a model training and content recall method, medium, device and computing device to solve the problem of low accuracy of recall caused by poor generalization of the recall model.

[0007] In a first aspect of the embodiments of the present disclosure, a model training method is provided, comprising:

[0008] obtaining a plurality of groups of first samples and a plurality of groups of second samples;

[0009] training a first factorization machine model according to the first samples to obtain a trained first factorization machine model, the first samples including at least one user portrait of a sample user, a content identifier of a sample content and corresponding sample click information, the sample click information being used to indicate that the sample user has clicked the sample content or has not clicked the sample content;

[0010] The second factorization machine model is trained according to the second sample, to obtain a trained second factorization machine model, wherein the second sample comprises a user identifier of the sample user, at least one content label of the sample content, and the sample click information.

[0011] In a second aspect of the embodiments of the present disclosure, a content recall method is provided, comprising:

[0012] A target user portrait vector of a target user and target content identifier vectors of a plurality of target contents are obtained according to the first factorization machine model.

[0013] A target user identifier vector of the target user and target content label vectors of the plurality of target contents are obtained according to the second factorization machine model.

[0014] A recall content is determined from the plurality of target contents according to the target user portrait vector, the target content identifier vectors, the target user identifier vector, and the target content label vectors.

[0015] In a third aspect of the embodiments of the present disclosure, a model training apparatus is provided, comprising:

[0016] An obtaining module is configured to obtain a plurality of first samples and a plurality of second samples.

[0017] A first training module is configured to train a first factorization machine model according to the first sample, to obtain a trained first factorization machine model, wherein the first sample comprises at least one user portrait of a sample user, a content identifier of a sample content, and corresponding sample click information, wherein the sample click information is used to indicate that the sample user has clicked the sample content, or has not clicked the sample content.

[0018] A second training module is configured to train a second factorization machine model according to the second sample, to obtain a trained second factorization machine model, wherein the second sample comprises a user identifier of the sample user, at least one content label of the sample content, and the sample click information.

[0019] In a fourth aspect of the embodiments of the present disclosure, a content recall apparatus is provided, comprising:

[0020] A first processing module is configured to obtain a target user portrait vector of a target user and target content identifier vectors of a plurality of target contents according to a first factorization machine model.

[0021] A second processing module is configured to obtain a target user identifier vector of the target user and target content label vectors of the plurality of target contents according to a second factorization machine model.

[0022] The recall module is used to determine the recall content from the plurality of target contents based on the target user profile vector, the target content identifier vector, the target user identifier vector, and the target content tag vector.

[0023] In a fifth aspect of the present disclosure, a computing device is provided, comprising: at least one processor and a memory;

[0024] The memory stores computer-executed instructions;

[0025] The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform a model training method as described in any of the first aspects, or causing the at least one processor to perform a content retrieval method as described in any of the second aspects.

[0026] In a sixth aspect of the present disclosure, a computer-readable storage medium is provided, wherein computer-executable instructions are stored, which, when executed by a processor, implement the model training method as described in any of the first aspects, or implement the content retrieval method as described in any of the second aspects.

[0027] In a seventh aspect of the present disclosure, a computer program product is provided, the computer program product comprising a computer program; when the computer program is executed, it implements the model training method according to any one of the first aspects, or when the computer program is executed, it implements the content recall method according to any one of the second aspects.

[0028] The model training and content retrieval method, medium, apparatus, and computing device provided in this disclosure first acquire multiple sets of first samples and multiple sets of second samples. Then, a first factor decomposition machine model is trained based on the first samples to obtain a trained first factor decomposition machine model. A second factor decomposition machine model is then trained based on the second samples to obtain a trained second factor decomposition machine model. The trained first and second factor decomposition machine models can be used in a recommendation system to retrieve content for users. Since the first factor decomposition machine model is trained using the first samples, which include user profiles and content identifiers, and the second factor decomposition machine model is trained using the second samples, which include user identifiers and sample content, compared to training a model solely using identifier-based features, the model in this disclosure combines tag-based features such as user profiles and content tags for training. This results in a model with better generalization ability and better adaptability to new samples, thereby improving the accuracy of content retrieval. Attached Figure Description

[0029] The above and other objects, features and advantages of the present disclosure exemplary embodiments will be more apparent from the following detailed description read in conjunction with the accompanying drawings, in which:

[0030] Figure 1 A schematic diagram of an application scenario provided by an embodiment of the present disclosure is shown in FIG. 1.

[0031] Figure 2 A schematic diagram of a model training method provided by an embodiment of the present disclosure is shown in FIG. 2.

[0032] Figure 3 A schematic diagram of a model training process provided by an embodiment of the present disclosure is shown in FIG. 3.

[0033] Figure 4 A schematic diagram of a content recall method provided by an embodiment of the present disclosure is shown in FIG. 4.

[0034] Figure 5 A schematic diagram of content recall provided by an embodiment of the present disclosure is shown in FIG. 5.

[0035] Figure 6 A schematic diagram of content recommendation provided by an embodiment of the present disclosure is shown in FIG. 6.

[0036] Figure 7 A schematic diagram of a storage medium provided by an embodiment of the present disclosure is shown in FIG. 7.

[0037] Figure 8 A schematic diagram of a model training apparatus provided by an embodiment of the present disclosure is shown in FIG. 8.

[0038] Figure 9 A schematic diagram of a content recall apparatus provided by an embodiment of the present disclosure is shown in FIG. 9.

[0039] Figure 10 A schematic diagram of a computing device provided by an embodiment of the present disclosure is shown in FIG. 10.

[0040] In the drawings, identical or corresponding reference signs indicate identical or corresponding parts. DETAILED DESCRIPTION

[0041] The principles and spirits of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and implement the present disclosure, and do not limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0042] Those skilled in the art understand that the embodiments of the present disclosure can be implemented as a system, a device, an apparatus, a method or a computer program product. Therefore, the present disclosure can be embodied as a complete hardware, a complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0043] According to embodiments of the present disclosure, a model training and content recall method, medium, device and computing device are provided.

[0044] In this document, it should be understood that any number of elements in the drawings are used for illustration only and not limitation, and any naming is only for differentiation and does not have any limiting meaning.

[0045] First, the basic concepts involved in the present disclosure are introduced.

[0046] ROC: Receiver Operating Characteristic, Receiver Operating Characteristic. According to the prediction result of the learner, the samples are sorted in descending order, and then the prediction result of each sample is predicted as a positive example. Each time, calculate the false positive rate (False Positive Rate, FPR) and the true positive rate (True Positive Rate, TPR), and then plot the two values as the horizontal and vertical coordinates, respectively, to obtain the corresponding ROC curve. The horizontal axis of the ROC curve is the "false positive rate", also known as the "false positive rate", which refers to the proportion of negative samples that are incorrectly predicted as positive examples; the vertical axis is the "true positive rate", also known as the "true positive rate", which refers to the proportion of positive samples that are correctly predicted as positive examples.

[0047] AUC: Area Under the Curve, the area under the ROC curve, used to indicate the classification ability of the learner, the value of AUC is positively correlated with the classification ability of the learner, the better the classification ability, the larger the value of AUC.

[0048] Factorization Machine (FM): a machine learning algorithm based on matrix factorization, which is usually used to train a classifier and is usually used to build a click rate estimation model in a recommendation system. Factorization machine has good learning ability for sparse data.

[0049] Hyperparameter: a parameter set before starting the learning process in machine learning, which is mainly related to the hyperparameters of the factorization machine model in the embodiments of the present disclosure. Unlike the parameters obtained during model training, hyperparameters need to be set before factorization machine model training, and hyperparameters remain unchanged during the entire model training process.

[0050] Generalization: refers to the ability of a model to make accurate predictions after training, also known as the ability to adapt to new samples.

[0051] The principles and spirits of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure. SUMMARY

[0053] The emergence and popularity of the Internet have brought a large amount of information to users, meeting the users' demand for information in the information age. However, with the rapid development of the Internet, the amount of information on the Internet has also increased significantly. In the face of a large amount of information, it becomes more difficult for users to obtain the information they need, and the efficiency of using information is actually reduced, which is the problem of information overload.

[0054] To solve the problem of information overload, there are two solutions. One is to search for information through an information retrieval system such as a search engine. Search engines play an extremely important role in helping users obtain network information. However, different search engines retrieve information in the same way for the same search keywords, while information dissemination is diversified, and user demand for information is also diversified. The information retrieval system represented by the search engine cannot meet the personalized needs of different users when searching for information.

[0055] The other solution is to search for information through a personalized recommendation system. The personalized recommendation system is a recommendation system that recommends information or products of interest to users according to their needs and interests. The personalized recommendation system performs personalized computing by obtaining users' interest preferences, and mines users' interest points to achieve accurate personalized recommendation.

[0056] The personalized recommendation system can solve the problem of information overload on the basis of meeting the personalized needs of users, and filter out a small amount of valuable information from a large amount of information. The current personalized recommendation system mainly includes two stages of recall and sorting. The recall stage mainly selects a batch of recommended content as candidate content from a large amount of content, and the sorting stage sorts the recommended content according to certain rules, and finally selects the content to be recommended to the user according to the sorting result for recommendation.

[0057] Since the content finally recommended to the user is selected from the recalled content, the accuracy of the recall is crucial to the final recommendation. The recall method in the current recommendation system is mainly to realize the recall based on a factor decomposition machine model. The factor decomposition machine model can be trained by a user id and a content id. Specifically, first, a training sample is obtained, which can include a user id, a content id and corresponding label information, the label information is used to label whether the user corresponding to the user id is interested in the content corresponding to the content id. Through the above training sample, the factor decomposition machine model can be trained, and after the training is completed, the content interested by each user can be selected according to the model.

[0058] The present inventors find that since the user identifiers (user ids) of different users are different, the content identifiers (content ids) of different contents are also different, the above factor decomposition machine model is trained by the user id and the content id, the generalization of the model is insufficient, and the adaptability to new samples is poor, that is, the recall accuracy of the model for new users and new contents will be reduced. Based on this, the embodiments of the present disclosure provide a model training and content recall method to solve the above technical problems.

[0059] After introducing the basic principles of the present disclosure, the various non-limiting embodiments of the present disclosure will be specifically introduced below.

[0060] OVERVIEW OF APPLICATION SCENARIOS

[0061] Firstly, the application scenario of the embodiments of the present disclosure is introduced with reference to Figure 1

[0062] Figure 1 The application scenario provided by the embodiments of the present disclosure is shown in Figure 1 The client 11 and the server 12 are connected through wired or wireless connection.

[0063] The user can view the content recommended for the user through the client 12, since personalized recommendation needs to be realized for different users, therefore, the relevant information of the user needs to be obtained through the client 12, for example, including the user identifier of the user, the user portrait of the user and the like. The relevant information of the user can be sent by the client 11 to the server 12.

[0064] After receiving the relevant information of the user, the server 12 can perform personalized recommendation according to the relevant information of the multiple contents and the relevant information of the user, wherein the process of personalized recommendation is completed based on a personalized recommendation system.

[0065] ​The personalized recommendation system includes two stages of recall and ranking, and the embodiment of the present disclosure mainly relates to the recall stage. In the recall stage, the server 12 determines the recalled content among the multiple contents based on the pre-trained recall model, and then selects part of the content in the recalled content to recommend to the client 12 in the ranking stage. The pre-trained recall model can be trained by the server 12, or trained by other devices and then sent to the server 12, and the embodiment of the present disclosure does not limit this.

[0066] Exemplary Method

[0067] The model training method according to the exemplary embodiments of the present disclosure will be described below in combination with an application scenario of Figure 1 Figure 2 It should be noted that the above-mentioned application scenario is only shown for the purpose of facilitating the understanding of the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0068] In the technical solution of the present disclosure, the collection of user behavior data and other information is based on the full authorization of the user, and the collection, storage, use, processing, transmission, provision and disclosure of these information comply with the relevant legal regulations and do not violate public order and good customs.

[0069] Figure 2 The flowchart of the model training method provided by the embodiments of the present disclosure is shown in Figure 2 The method can include the following steps:

[0070] S21, obtaining a plurality of first samples and a plurality of second samples.

[0071] The execution subject of the embodiments of the present disclosure can be a server, or an electronic device integrated with the functions of the server. Before executing the model training, a plurality of first samples and a plurality of second samples are first obtained, wherein the first samples are used to train the first factorization machine model, and the second samples are used to train the second factorization machine model.

[0072] S22, training the first factorization machine model according to the first samples to obtain a trained first factorization machine model, wherein the first samples include at least one user portrait of a sample user, a content identifier of a sample content and corresponding sample click information, and the sample click information is used to indicate that the sample user has clicked the sample content or has not clicked the sample content.

[0073] The user portrait is a label abstracted from the specific information of the user, and these labels can concretize the image of the user. The at least one user portrait in the first sample is at least one label abstracted from the specific information of the sample user.​

[0074] The content identifier is a label used to distinguish different contents. The content identifier can be a number, a letter, a symbol, a combination, or the like. Different contents have different content identifiers, and each content identifier is unique.

[0075] The sample click information is used to indicate whether the sample user corresponding to the user portrait clicks the sample content corresponding to the content identifier. For example, the sample click information can be 0 or 1, where 0 indicates that the sample user has not clicked the sample content, and 1 indicates that the sample user has clicked the sample content.

[0076] The first sample is used to train the first factorization machine model. The user portrait and the content identifier are used as inputs to the first factorization machine model. The first factorization machine model processes the user portrait and the content identifier. The sample click information is the labeled information. The first factorization machine model can be trained according to the sample click information and the output of the first factorization machine model. After the first factorization machine model is trained by multiple sets of first samples, a trained first factorization machine model can be obtained.

[0077] S23, training the second factorization machine model according to the second sample to obtain a trained second factorization machine model, the second sample including the user identifier of the sample user, at least one content label of the sample content, and sample click information.

[0078] The user identifier is a label used to distinguish different users. The user identifier can be a number, a letter, a symbol, a combination, or the like. Different users have different user identifiers, and each user identifier is unique.

[0079] The content label is a label obtained by abstracting the corresponding content. These labels can be used to reflect the relevant information of the corresponding content. The at least one content label in the second sample is at least one label obtained according to the sample content.

[0080] The sample click information is used to indicate whether the sample user corresponding to the user identifier clicks the sample content corresponding to the content label. For example, the sample click information can be 0 or 1, where 0 indicates that the sample user has not clicked the sample content, and 1 indicates that the sample user has clicked the sample content.

[0081] The second sample is used to train the second factorization machine model. The user identifier and the content label are used as inputs to the second factorization machine model. The second factorization machine model processes the user identifier and the content label. The sample click information is the labeled information. The second factorization machine model can be trained according to the sample click information and the output of the second factorization machine model. After the second factorization machine model is trained by multiple sets of second samples, a trained second factorization machine model can be obtained.

[0082] After the first factorization machine model and the second factorization machine model are trained, the first factorization machine model and the second factorization machine model can be used to form a recall system for recalling content for a user.

[0083] The model training method provided in the embodiments of the present disclosure first acquires a plurality of first samples and a plurality of second samples, then trains the first factorization machine model according to the first samples to obtain the trained first factorization machine model, and trains the second factorization machine model according to the second samples to obtain the trained second factorization machine model. The first factorization machine model and the second factorization machine model obtained through training can be used in a recommendation system to recall content for a user. Since the first factorization machine model is trained by the first samples, the first samples include user portraits and content identifiers, and the second factorization machine model is trained by the second samples, the second samples include user identifiers and sample content, compared with training a model by only using identifier features, the model in the embodiments of the present disclosure is trained by combining user portraits and content labels and other label features, so that the trained model has better generalization and better adaptability to new samples, thereby improving the accuracy of recalling content.

[0084] In the above embodiments, the training of the first factorization machine model and the second factorization machine model is introduced, and the scheme of the present disclosure will be described in detail in combination with the drawings.

[0085] Before training the model, training samples, i.e. a plurality of first samples and a plurality of second samples, need to be acquired. Specifically, a plurality of initial samples are first acquired, and the initial samples include user identifiers of sample users, content identifiers, and sample click information.

[0086] Figure 3 The model training process provided in the embodiments of the present disclosure is shown in FIG. 1, and the initial samples can be acquired by acquiring user behavior data of sample users. The user behavior data indicates whether the sample users have clicked to view sample content. For a sample user, the sample user may have browsed one or more sample contents, and thus the user identifier of the sample user, the content identifier of any sample content, and the sample click information of the sample user on the sample content can constitute an initial sample. Figure 3 As shown in FIG. 2, the user behavior data of one or more sample users can be acquired, and then the click information can be extracted from the user behavior data, where the click information indicates whether the sample user has clicked a certain sample content. For the user behavior data of any sample user, the user identifier of the sample user, the content identifier of the sample content, and the corresponding click information can be acquired from the user behavior data, thereby obtaining the initial sample. Figure 3 ​

[0087] After obtaining the plurality of sets of initial samples, at least one user portrait of the sample user and at least one content label of the sample content can be obtained, and the user portrait and the content label can be obtained by using a pre-trained model.

[0088] For the user portrait, a user portrait model can be used to obtain the user portrait, and the user portrait model can be pre-trained. For example, user behavior data and a sample user portrait can be obtained, and the sample user portrait is the labeled information of the user behavior data. The user behavior data is input into the user portrait model to obtain a user portrait output by the user portrait model, and then the parameters of the user portrait model are adjusted according to the difference between the user portrait output by the user portrait model and the sample user portrait. For any set of user behavior data and sample user portrait, the user portrait model can be trained in the above manner, and finally a trained user portrait model is obtained. The trained user portrait model has the ability to output a user portrait according to user behavior data. Therefore, the user behavior data of the sample user can be input into the trained user portrait model to obtain at least one user portrait of the sample user.

[0089] For the content label, a content label model can be used to obtain the content label, and the content label model can be pre-trained. For example, sample content and a sample content label can be obtained, and the sample content label is the labeled information of the sample content. The sample content is input into the content label model to obtain a content label output by the content label model, and then the parameters of the content label model are adjusted according to the difference between the content label output by the content label model and the sample content label. For any set of sample content and sample content label, the content label model can be trained in the above manner, and finally a trained content label model is obtained. The trained content label model has the ability to output a content label according to content. Therefore, the sample content can be input into the trained content label model to obtain at least one content label of the sample content.

[0090] After obtaining at least one user portrait of the sample user and at least one content label of the sample content, a first sample can be obtained according to the at least one user portrait, the content label, and the sample click information, and a second sample can be obtained according to the user identifier, the at least one content label, and the sample click information.

[0091] After obtaining the first sample, the first sample can be divided to obtain a first training sample and a first verification sample, and the number ratio of the first training sample and the first verification sample can be set as needed, and the present disclosure does not limit this. The first training sample is used to train the first factorization machine model, and the first verification sample is used to verify whether the trained first factorization machine model meets the requirements.

[0092] The first factorization machine model and the second factorization machine model both belong to factorization machine models, and the training manners of the two models are similar, but the training samples are different. Before introducing the training processes of the two models, the factorization machine model is introduced first.

[0093] In machine learning, a common linear model is:

[0094] y = ω1x1 + ω2x2 +... + ω n x n , (1)

[0095] where y is the output of the model, x i is the feature input into the model, and ω i is the parameter of the model, that is, the weight of the feature x i . The disadvantage of the linear model is that it cannot capture the mutual relationship between the features. In order to solve this problem, polynomial regression can be introduced as follows:

[0096]

[0097] However, since the weight ω j,i between each two features needs to be learned, the number of parameters to be learned is n(n-1) / 2, and n is the number of features. The value of n is often large, and the model is particularly complex, and the actual data is usually sparse, which makes it difficult for the model to learn the parameters.

[0098] Therefore, the factorization machine can be derived from matrix factorization, and based on the matrix factorization idea, a sparse matrix is decomposed into the product of two low-order matrices. In the embodiment of the disclosure, the following factorization machine model is designed for user class features and content class features:

[0099]

[0100] where y is the output of the model, x i is the feature input into the model, ω i is the weight parameter learned by the feature, and p is the number of features of each sample.

[0101] Let x u represent the user feature, ω u and represent the weight parameter learned by the user feature, x d represent the content feature, ω d and represent the weight parameter learned by the content feature, then formula (3) can be expressed in the following form:

[0102]

[0103] Equation (4) is the factorization machine model in the embodiments of this disclosure, where y is the output of the factorization machine model. For the j-th user feature, p u For the number of user features, For the j-th content feature, p d ω0 represents the number of content features. and These are all parameters of the factorization machine model.

[0104] make Equation (4) can then be expressed in the form of equation (5):

[0105]

[0106] make Equation (5) can be further expressed as equation (6):

[0107]

[0108] make Then, according to equation (6), we can further obtain:

[0109]

[0110] Where ω0 is the parameter of the factorization machine model, and ω0 is a constant when the factorization machine model has been trained. For user feature vectors, This is the content feature vector.

[0111] The structure of the factorization machine model in the embodiments of this disclosure has been described from equation (1) to equation (7). The first factorization machine model and the second factorization machine model in the embodiments of this disclosure will be described below based on the factorization machine model in the above embodiments.

[0112] Based on the structure of the factorization machine model shown in equation (4), the structure of the first factorization machine model can be obtained as follows:

[0113]

[0114] Where y1 is the output of the first factorization machine model, ω 01 , and These are all parameters of the first factorization machine model. Since the first factorization machine model is trained using the first training samples, which include at least one user profile of the sample user and a content identifier of the sample content, the user feature in the first factorization machine model is the user profile, and the content feature is the content identifier. For the j-th group of samples in the first training samples, is the jth user portrait, p u1 is the number of user portraits, is the jth content identifier, p d1 is the number of content identifiers.

[0115] Let Then, according to formula (8), the following can be further obtained:

[0116]

[0117] When training the first factorization machine model, for any one set of first training samples, the corresponding user portrait and content identifier can be input into the first factorization machine model, and according to formula (8), the output y1 of the first factorization machine model can be obtained, so as to obtain the user portrait vector and the content identifier vector output by the first factorization machine model.

[0118] For example, one possible implementation is that according to y1 and formula (9), the inner product of the user portrait vector and the content identifier vector can be obtained. After obtaining the inner product of the user portrait vector and the content identifier vector , the inner product can be mapped using a sigmoid function to the range [0, 1], and the result after mapping is the first click information, which is a value in the range [0, 1].

[0119] Then, the parameters of the first factorization machine model are adjusted according to the first click information and the sample click information. The sample click information is used to indicate that the sample user has clicked the sample content or has not clicked the sample content. If the sample user has clicked the sample content, the sample click information takes the value 1, and if the sample user has not clicked the sample content, the sample click information takes the value 0. Then, according to the difference between the first click information and the sample click information, the parameters (such as ω 01 、 and in formula (8)) in the first factorization machine model can be adjusted. For any one set of first training samples, the first factorization machine model can be trained according to the above implementation, and finally the trained first factorization machine model is obtained. After the first factorization machine model is trained, the values of the parameters in the first factorization machine model are determined, so that after the user portrait and the content identifier are input into the first factorization machine model, the output of the first factorization machine model can be obtained according to the parameters in the first factorization machine model and formula (8). At the same time, according to and The corresponding user profile vector and content identifier vector can be obtained.

[0120] like Figure 3 As shown, after training the first factorization machine model based on the first training sample, the first validation sample can be used to verify whether the first factorization machine model meets the requirements. One possible implementation is to input the user profile and content identifier in the first validation sample into the first factorization machine model to obtain the click information output by the first factorization machine model. Specifically, based on the user profile and content identifier in the first validation sample, and the various parameters in the trained first factorization machine model, the inner product of the corresponding user profile vector and content identifier vector can be obtained by combining Equation (8). Then, the inner product is mapped by the sigmoid function to obtain the click information output by the first factorization machine model.

[0121] Then, based on the click information output by the first factorization machine model and the corresponding sample click information in the first validation sample, the area under the first ROC curve of the first factorization machine model is obtained. After obtaining the area under the first ROC curve, the first validation result can be determined based on the area under the first ROC curve. The first validation result is used to indicate whether the trained first factorization machine model meets the requirements.

[0122] like Figure 3 As shown, if the first verification result indicates that the trained first factorization machine model does not meet the requirements, the hyperparameters of the first factorization machine model can be updated. The updated hyperparameters may include, for example, the degree of the first factorization machine model, or the learning rate, etc. Then, the first factorization machine model is retrained based on the updated hyperparameters and the first training samples. The training method is as described in the above embodiments and will not be repeated here. By optimizing the hyperparameters of the first factorization machine model, the learning performance and effectiveness of the first factorization machine model can be improved. If the first verification result indicates that the trained first factorization machine model meets the requirements, training can be stopped, and the obtained first factorization machine model can be used to subsequently output user profile vectors and content identifier vectors based on user profiles and content identifiers.

[0123] Based on the structure of the factorization machine model shown in equation (4), the structure of the second factorization machine model can be obtained as follows:

[0124]

[0125] Where y2 is the output of the second factorization machine model, ω 02 , and are parameters of the second factorization machine model. Since the second factorization machine model is trained according to the second training samples, the second training samples include user identifiers of sample users and at least one content label of sample contents, the user feature in the second factorization machine model is the user identifier, and the content feature is the content label. For the jth group of samples in the second training samples, is the jth user identifier, p u2 is the number of user identifiers, is the jth content label, p d2 is the number of content labels.

[0126] Let According to equation (10), the following can be further obtained:

[0127]

[0128] When training the second factorization machine model, for any group of second training samples, the corresponding user identifier and content label can be input into the second factorization machine model, and according to equation (10), the output y2 of the second factorization machine model can be obtained, thereby obtaining the user identifier vector and the content label vector output by the second factorization machine model.

[0129] For example, one possible implementation is that according to y2 and equation (11), the inner product of the user identifier vector and the content label vector can be obtained. After obtaining the inner product of the user identifier vector and the content label vector , the sigmoid function can be used to map the inner product to the range of [0, 1], and the result after mapping is the second click information, which is a value in the range of [0, 1].

[0130] Then, the parameters of the second factorization machine model are adjusted according to the second click information and the sample click information. The sample click information is used to indicate that the sample user has clicked the sample content or has not clicked the sample content. If the sample user has clicked the sample content, the sample click information takes the value of 1, and if the sample user has not clicked the sample content, the sample click information takes the value of 0. Then, according to the difference between the second click information and the sample click information, the parameters in the second factorization machine model (for example, including ω 02 、 and Adjustments are made accordingly. For any set of second training samples, the second factorization machine model can be trained according to the above implementation method to finally obtain the trained second factorization machine model. After the second factorization machine model is trained, the values ​​of each parameter in the second factorization machine model are determined. Therefore, after inputting the user identifier and content tag into the second factorization machine model, the output of the second factorization machine model can be obtained according to the parameters of the second factorization machine model combined with equation (10). At the same time, according to and The corresponding user identifier vector and content tag vector can be obtained.

[0131] like Figure 3 As shown, after training the second factorization machine model based on the second training samples, the second validation samples can be used to verify whether the second factorization machine model meets the requirements. One possible implementation is to input the user identifier and content label from the second validation samples into the second factorization machine model to obtain the click information output by the second factorization machine model. Specifically, based on the user identifier and content label from the second validation samples, and the parameters in the trained second factorization machine model, the inner product of the corresponding user identifier vector and content label vector can be obtained by combining Equation (10). Then, the inner product is mapped using the sigmoid function to obtain the click information output by the second factorization machine model.

[0132] Then, based on the click information output by the second factorization machine model and the corresponding sample click information in the second validation samples, the area under the second ROC curve of the second factorization machine model is obtained. After obtaining the area under the second ROC curve, the second validation result can be determined based on the area under the second ROC curve. The second validation result is used to indicate whether the trained second factorization machine model meets the requirements.

[0133] like Figure 3 As shown, if the second verification result indicates that the trained second factorization machine model does not meet the requirements, the hyperparameters of the second factorization machine model can be updated. Updated hyperparameters may include, for example, the degree of the second factorization machine model, the learning rate of the second factorization machine model, etc. Then, the second factorization machine model is retrained based on the updated hyperparameters and the second training samples. The training method is as described in the above embodiments and will not be repeated here. By optimizing the hyperparameters of the second factorization machine model, the learning performance and effectiveness of the second factorization machine model can be improved. If the second verification result indicates that the trained second factorization machine model meets the requirements, training can be stopped, and the obtained second factorization machine model can be used to subsequently output user identifier vectors and content tag vectors based on user identifiers and content tags.

[0134] In the above embodiments, the first factorization machine model and the second factorization machine model were introduced. The following will describe, with reference to the accompanying drawings, how to perform content retrieval based on the first factorization machine model and the second factorization machine model.

[0135] Figure 4 This is a flowchart illustrating the content recall method provided in the embodiments of this disclosure, such as... Figure 4 As shown, the method may include:

[0136] S41, obtain the target user profile vector of the target user and the target content identifier vector of multiple target contents based on the first factor decomposition machine model.

[0137] Figure 5 This is a schematic diagram illustrating the content recall provided in the embodiments of this disclosure, such as... Figure 5 As shown, the target user profile vector and target content identifier vector can be obtained based on the first factorization machine model. The first factorization machine model is based on... Figure 3 The model trained by the method described in the examples.

[0138] Specifically, when it is necessary to recommend content to a target user, the first step is to obtain the target user profile and target content identifiers for multiple target content items. The target user profile can be obtained through a pre-trained user profile model. Then, based on the target user profile and its corresponding parameters in the first factorization machine model, the target user profile vector is obtained; and based on the target content identifiers and their corresponding parameters in the first factorization machine model, the target content identifier vector is obtained.

[0139] S42, obtain the target user identifier vector of the target user and the target content tag vector of multiple target contents according to the second factorization machine model.

[0140] like Figure 5 As shown, the target user identifier vector and target content tag vector can be obtained based on the second factorization machine model. The second factorization machine model is based on... Figure 3 The model trained by the method described in the examples.

[0141] Specifically, firstly, the target user identifier and multiple target content tags are obtained. The target content tags can be obtained through a pre-trained content tag model. Then, based on the target user identifier and its corresponding parameters in the second factorization machine model, the target user identifier vector is obtained; and based on the target content tags and their corresponding parameters in the second factorization machine model, the target content tag vector is obtained.

[0142] S43, determine the recall content among multiple target contents based on the target user profile vector, target content identifier vector, target user identifier vector, and target content tag vector.

[0143] like Figure 5 As shown, after obtaining the target user profile vector, target content identifier vector, target user identifier vector, and target content tag vector, the target user profile vector and target user identifier vector are concatenated to obtain the target user vector of the target user; for any content among multiple content items, the target content identifier vector and target content tag vector are concatenated to obtain the target content vector of the content.

[0144] Then, based on the target user vector and the target content vector of multiple content, the recall content is determined among the multiple content. Figure 6 This is a schematic diagram illustrating the content recommendations provided for embodiments of this disclosure, such as... Figure 6 As shown, set 61 includes multiple target contents, namely articles 1-10. For user A, after obtaining user A's target user identifier and target user profile, as well as the target content tags and target content identifiers of each article, the target user vector of user A and the target content vector of each article are obtained based on the first factorization machine model and the second factorization machine model.

[0145] Then, the inner product of the target user vector and the target content vector of each article is performed to obtain the inner product score of each target content for the target user. Faiss can be used to build an index of the target content vector, and Faiss can be used to retrieve the N highest-scoring target content articles using the target user vector. For example, in... Figure 6 In the process, the inner product score of articles 1-10 for user A is calculated. Based on the inner product score, articles 1-10 can be sorted, such as... Figure 6 The diagram illustrates the sorting results. If the pre-set number of recalled content items is 3, then Article 1, Article 3, and Article 6, which rank higher in the sorting results, can be used as the recalled content, as shown in set 62. After the recall phase, the recalled content can be sorted according to a certain sorting method to determine the final content recommended to user A. For example, in interface 63, Article 3 out of Article 1, Article 3, and Article 6 can be recommended to user A.

[0146] The content recall method provided by the embodiments of the present disclosure first acquires a target user portrait vector of a target user and a target content identifier vector of a plurality of target contents according to a first factor decomposition machine model, then acquires a target user identifier vector of the target user and a target content label vector of the plurality of target contents according to a second factor decomposition machine model, and determines a recall content in the plurality of target contents according to the target user portrait vector, the target content identifier vector, the target user identifier vector and the target content label vector. Since the first factor decomposition machine model is trained by a user portrait of a sample user and a content identifier of a sample content, and the second factor decomposition machine model is trained by a user identifier of a sample user and a content label of a sample content, the recall of the content is comprehensively performed in combination with the identifier type feature and the label type feature, compared with a model trained only according to the identifier type feature and a model trained only according to the label type feature, the model of the embodiments of the present disclosure has better generalization, so as to improve the accuracy of the recall and realize more accurate content recommendation.

[0147] Exemplary Medium

[0148] After introducing the method of the exemplary embodiments of the present disclosure, next, reference is made to Figure 7 The storage medium of the exemplary embodiments of the present disclosure is described.

[0149] Figure 7 The storage medium of the exemplary embodiments of the present disclosure is described. Figure 7 As shown in FIG. 7, the storage medium 70 stores a program product for implementing the above method according to the embodiments of the present disclosure, which can adopt a portable compact disc read-only memory (CD-ROM) and include program codes, and can run on a terminal device such as a personal computer. However, the program product of the present disclosure is not limited thereto.

[0150] The program product can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0151] A non-transitory computer readable medium can include data that causes a machine to operate in a predetermined manner. Alternatively or additionally, a non-transitory computer readable medium can include instructions or software that cause a machine to operate in a predetermined manner. Other examples and implementations are within the scope of the disclosure and appended claims. For example, implementation of the examples can produce a machine that operates "online" (e.g., on the internet) or "off-line" (e.g., functioning independently or in an offline computing environment). Such machines can be connected (e.g., via wired or wireless communication means) to other machines. In an online implementation, the machine can communicate with a server on a reliable internet connection.

[0152] The program code for carrying out operations of the present disclosure can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++, or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on a remote computing device, or entirely on the remote computing device or server. In the latter scenario, the remote computing device can be connected to the user's computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0153] Exemplary Apparatus

[0154] Having introduced the medium of the exemplary embodiments of the present disclosure, next, the method of the exemplary embodiments of the present disclosure will be described with reference to Figure 8 and Figure 9 The model training apparatus and the content recall apparatus of the exemplary embodiments of the present disclosure are described for implementing the method in any of the method embodiments described above, and have similar implementation principles and technical effects, which will not be described here.

[0155] Figure 8 The structural schematic diagram of the model training apparatus provided by the embodiments of the present disclosure is shown in FIG. 8, which includes: Figure 8

[0156] The acquisition module 81 is configured to acquire a plurality of groups of first samples and a plurality of groups of second samples.

[0157] The first training module 82 is configured to train a first factorization machine model according to the first samples to obtain a trained first factorization machine model, wherein the first samples include at least one user portrait of a sample user, a content identifier of a sample content, and corresponding sample click information, and the sample click information is used to indicate that the sample user has clicked the sample content or has not clicked the sample content.

[0158] The second training module 83 is configured to train a second factorization machine model according to the second samples to obtain a trained second factorization machine model, wherein the second samples include a user identifier of the sample user, at least one content label of the sample content, and the sample click information.

[0159] ​In one possible implementation, the acquisition module 81 is specifically used for:

[0160] Obtain multiple sets of initial samples, wherein the initial samples include the user identifier, the content identifier, and the sample click information;

[0161] Obtain at least one user profile of the sample user and at least one content tag of the sample content;

[0162] The first sample is obtained based on the at least one user profile, the content identifier, and the sample click information;

[0163] The second sample is obtained based on the user identifier, the at least one content tag, and the sample click information.

[0164] In one possible implementation, the first sample includes a first training sample; the first training module 82 is specifically used for:

[0165] Input the user profile and content identifier from the first training sample into the first factorization machine model to obtain the user profile vector and content identifier vector output by the first factorization machine model.

[0166] The first click information is obtained based on the user profile vector and the content identifier vector;

[0167] The parameters of the first factorization machine model are adjusted based on the first click information and the sample click information to obtain the trained first factorization machine model.

[0168] In one possible implementation, the first sample further includes a first verification sample; the first training module 82 is further configured to:

[0169] The user profile and content identifier in the first verification sample are input into the first factorization machine model to obtain the click information output by the first factorization machine model.

[0170] Based on the click information output by the first factorization machine model and the corresponding sample click information, the area under the ROC curve of the first receiver operating characteristic of the first factorization machine model is obtained.

[0171] The first verification result is determined based on the area under the first ROC curve. The first verification result is used to indicate whether the trained first factorization machine model meets the requirements.

[0172] In one possible implementation, if the first verification result indicates that the trained first factorization machine model does not meet the requirements, the first training module 82 is further configured to:

[0173] update the hyperparameters of the first factorization machine model, and retrain the first factorization machine model according to the updated hyperparameters and the first training sample.

[0174] In a possible implementation, the second sample includes a second training sample; and the second training module 83 is specifically configured to:

[0175] input the user identifier and the content label in the second training sample into the second factorization machine model to obtain a user identifier vector and a content label vector output by the second factorization machine model;

[0176] obtain second click information according to the user identifier vector and the content label vector;

[0177] adjust parameters of the second factorization machine model according to the second click information and the sample click information, to obtain a trained second factorization machine model.

[0178] In a possible implementation, the second sample further includes a second verification sample; and the second training module 83 is further configured to:

[0179] input the user identifier and the content label in the second verification sample into the second factorization machine model to obtain click information output by the second factorization machine model;

[0180] obtain an area under a second ROC curve of the second factorization machine model according to the click information output by the second factorization machine model and the corresponding sample click information;

[0181] determine a second verification result according to the area under the second ROC curve, the second verification result being used to indicate whether the trained second factorization machine model meets a requirement.

[0182] In a possible implementation, if the second verification result indicates that the trained second factorization machine model does not meet the requirement, the second training module 83 is further configured to:

[0183] update the hyperparameters of the second factorization machine model, and retrain the second factorization machine model according to the updated hyperparameters and the second training sample.

[0184] The model training apparatus provided by the embodiments of the present disclosure can be used to execute the technical solutions of the method embodiments, and the implementation principles and technical effects are similar, which will not be described herein.

[0185] Figure 9 A structural diagram of the content recall apparatus provided by the embodiments of the present disclosure is shown in FIG. 1. Figure 9As shown, comprising:

[0186] The first processing module 91 is configured to obtain a target user portrait vector of a target user and target content identifier vectors of a plurality of target contents according to a first factorization machine model.

[0187] The second processing module 92 is configured to obtain a target user identifier vector of the target user and target content label vectors of the plurality of target contents according to a second factorization machine model.

[0188] The recall module 93 is configured to determine a recall content from the plurality of target contents according to the target user portrait vector, the target content identifier vectors, the target user identifier vector and the target content label vectors.

[0189] In a possible implementation, the recall module 93 is specifically configured to:

[0190] splicing the target user portrait vector and the target user identifier vector to obtain a target user vector of the target user;

[0191] splicing the target content identifier vectors and the target content label vectors of any content from the plurality of contents to obtain a target content vector of the content;

[0192] determining the recall content from the plurality of contents according to the target user vector and the target content vectors of the contents.

[0193] In a possible implementation, the recall module 93 is specifically configured to:

[0194] performing inner product processing on the target user vector and the target content vectors of the contents to obtain an inner product score of each content for the target user;

[0195] determining a preset number of target contents as the recall content from the plurality of contents according to the inner product scores.

[0196] In a possible implementation, the first factorization machine model includes parameters of user portraits and parameters of content identifiers; and the first processing module 91 is specifically configured to:

[0197] obtain a target user portrait of the target user and a target content identifier of the target content;

[0198] obtain the target user portrait vector according to the target user portrait and parameters corresponding to the target user portrait in the first factorization machine model;

[0199] According to the target user identifier and the parameter corresponding to the target user identifier in the second factorization machine model, a target user identifier vector is obtained.

[0200] In a possible implementation, the second factorization machine model includes parameters of each user identifier and parameters of each content label; and the second processing module 92 is specifically configured to:

[0201] obtain a target user identifier of the target user and a target content label of the target content;

[0202] According to the target user identifier and the parameter corresponding to the target user identifier in the second factorization machine model, a target user identifier vector is obtained.

[0203] According to the target content label and the parameter corresponding to the target content label in the second factorization machine model, a target content label vector is obtained.

[0204] The content recall apparatus provided by the embodiments of the present disclosure can be used to execute the technical solutions of the method embodiments, and has similar implementation principles and technical effects, which will not be described here.

[0205] Exemplary Computing Device

[0206] After introducing the method, medium and apparatus of the exemplary embodiments of the present disclosure, next, with reference to Figure 10 The computing device of the exemplary embodiments of the present disclosure is described.

[0207] Figure 10 The computing device 100 shown is merely an example, and should not bring any limitation to the functions and use range of the embodiments of the present disclosure.

[0208] Figure 10 The structural schematic diagram of the computing device provided by the embodiments of the present disclosure is shown in Figure 10 The computing device 100 is in the form of a general-purpose computing device. The components of the computing device 100 can include, but are not limited to, the above-mentioned at least one processing unit 101, the above-mentioned at least one storage unit 102, and a bus 103 connecting different system components, including the processing unit 101 and the storage unit 102.

[0209] The bus 103 includes a data bus, a control bus and an address bus.

[0210] The storage unit 102 can include a readable medium in the form of a volatile memory, such as a random access memory (RAM) 1021 and / or a cache memory 1022, and can further include a readable medium in the form of a non-volatile memory, such as a read-only memory (ROM) 1023.

[0211] Storage unit 102 may also include a program / utility 1025 having a set (at least one) program module 1024, such program module 1024 including but not limited to: operating system, one or more application programs, other program modules and program data, each of these examples or some combination of these may include an implementation of a network environment.

[0212] The computing device 100 can also communicate with one or more external devices 104 (e.g., keyboard, pointing device, etc.). This communication can be performed via the input / output (I / O) interface 105. Furthermore, the computing device 100 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via a network adapter 106. Figure 10 As shown, network adapter 106 communicates with other modules of computing device 100 via bus 103. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with computing device 100, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0213] It should be noted that although several units / modules or sub-units / modules of the model training apparatus and content retrieval apparatus have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0214] Furthermore, although the operations of the methods disclosed herein are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0215] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A model training method, comprising: Obtain multiple sets of first samples and multiple sets of second samples; The first sample includes at least one user profile of the sample user, a content identifier of the sample content, and corresponding sample click information. The sample click information is used to indicate whether the sample user clicked the sample content or not. The second sample includes the user identifier of the sample user, at least one content tag of the sample content, and the sample click information. The first user profile and the content identifier included in the first sample are input into the first factorization machine model to obtain the first click information; The parameters of the first factorization machine model are adjusted based on the first click information and the sample click information to obtain the trained first factorization machine model. The user identifier and at least one content tag included in the second sample are input into the second factorization machine model to obtain the second click information; The parameters of the second factorization machine model are adjusted based on the second click information and the sample click information to obtain the trained second factorization machine model. The first and second factorization machine models, after training, constitute a recall system.

2. The method according to claim 1, wherein obtaining multiple sets of first samples and multiple sets of second samples comprises: Obtain multiple sets of initial samples, wherein the initial samples include the user identifier, the content identifier, and the sample click information; Obtain at least one user profile of the sample user and at least one content tag of the sample content; The first sample is obtained based on the at least one user profile, the content identifier, and the sample click information; The second sample is obtained based on the user identifier, the at least one content tag, and the sample click information.

3. The method according to claim 1 or 2, wherein the first sample includes a first training sample; the step of inputting the at least one user profile and the content identifier included in the first sample into a first factorization machine model to obtain first click information includes: Input the user profile and content identifier from the first training sample into the first factorization machine model to obtain the user profile vector and content identifier vector output by the first factorization machine model. The first click information is obtained based on the user profile vector and the content identifier vector.

4. The method according to claim 3, wherein the first sample further includes a first verification sample; the method further includes: The user profile and content identifier in the first verification sample are input into the first factorization machine model to obtain the click information output by the first factorization machine model. Based on the click information output by the first factorization machine model and the corresponding sample click information, the area under the ROC curve of the first receiver operating characteristic of the first factorization machine model is obtained. The first validation result is determined based on the area under the first ROC curve. The first validation result is used to indicate whether the trained first factorization machine model meets the requirements.

5. The method according to claim 4, wherein if the first verification result indicates that the trained first factorization machine model does not meet the requirements, the method further comprises: Update the hyperparameters of the first factorization machine model, and retrain the first factorization machine model based on the updated hyperparameters and the first training samples.

6. The method according to claim 1 or 2, wherein the second sample includes a second training sample; the step of inputting the user identifier and the at least one content tag included in the second sample into a second factorization machine model to obtain second click information includes: Input the user identifier and content label from the second training sample into the second factorization machine model to obtain the user identifier vector and content label vector output by the second factorization machine model; The second click information is obtained based on the user identifier vector and the content tag vector.

7. The method according to claim 6, wherein the second sample further comprises a second verification sample; the method further comprises: The user identifier and content tag in the second verification sample are input into the second factorization machine model to obtain the click information output by the second factorization machine model; Based on the click information output by the second factorization machine model and the corresponding sample click information, the area under the second ROC curve of the second factorization machine model is obtained; The second verification result is determined based on the area under the second ROC curve. The second verification result is used to indicate whether the trained second factorization machine model meets the requirements.

8. The method according to claim 7, wherein if the second verification result indicates that the trained second factorization machine model does not meet the requirements, the method further comprises: Update the hyperparameters of the second factorization machine model, and retrain the second factorization machine model based on the updated hyperparameters and the second training samples.

9. A content recall method, comprising: Based on the first factor decomposition machine model, obtain the target user profile vector of the target user and the target content identifier vector of multiple target contents; The target user identifier vector of the target user and the target content tag vector of the multiple target contents are obtained according to the second factor decomposition machine model. Based on the target user profile vector, the target content identifier vector, the target user identifier vector, and the target content tag vector, the recall content is determined from the plurality of target contents; The first factorization machine model and the second factorization machine model are trained based on the model training method as described in any one of claims 1-8.

10. The method according to claim 9, wherein determining the recall content from the plurality of target contents based on the target user profile vector, the target content identifier vector, the target user identifier vector, and the target content tag vector comprises: The target user profile vector and the target user identifier vector are concatenated to obtain the target user vector of the target user. For any content among the multiple target contents, the target content identifier vector and the target content tag vector of the content are concatenated to obtain the target content vector of the content. The recall content is determined from the plurality of target contents based on the target user vector and the target content vector of each of the contents.

11. The method according to claim 10, wherein determining the recalled content among the plurality of content based on the target user vector and the target content vector of each of the content comprises: Perform an inner product operation on the target user vector and the target content vector of each content to obtain the inner product score of each content for the target user; Based on the inner product score, a preset number of target contents are determined from the plurality of contents as the recall contents.

12. The method according to any one of claims 9-11, wherein the first factorization machine model includes parameters for each user profile and parameters for each content identifier; the step of obtaining the target user profile vector of the target user and the target content identifier vector of multiple target contents according to the first factorization machine model includes: Obtain the target user profile of the target user and the target content identifier of the target content; The target user profile vector is obtained based on the target user profile and the parameters corresponding to the target user profile in the first factorization machine model. The target content identifier vector is obtained based on the target content identifier and the parameters corresponding to the target content identifier in the first factorization machine model.

13. The method according to any one of claims 9-11, wherein the second factorization machine model includes parameters for each user identifier and parameters for each content tag; The step of obtaining the target user identifier vector of the target user and the target content tag vector of the multiple target contents according to the second factorization machine model includes: Obtain the target user identifier of the target user and the target content tag of the target content; The target user identifier vector is obtained based on the target user identifier and the parameters corresponding to the target user identifier in the second factorization machine model; The target content tag vector is obtained based on the target content tag and the parameters corresponding to the target content tag in the second factorization machine model.

14. A model training device, comprising: The acquisition module is used to acquire multiple sets of first samples and multiple sets of second samples. The first sample includes at least one user profile of the sample user, a content identifier of the sample content, and corresponding sample click information. The sample click information is used to indicate whether the sample user clicked the sample content or not. The second sample includes the user identifier of the sample user, at least one content tag of the sample content, and the sample click information. The first training module is used to input the at least one user profile and the content identifier included in the first sample into the first factorization machine model to obtain the first click information; and to adjust the parameters of the first factorization machine model according to the first click information and the sample click information to obtain the trained first factorization machine model. The second training module is used to input the user identifier and at least one content tag included in the second sample into the second factorization machine model to obtain the second click information; adjust the parameters of the second factorization machine model according to the second click information and the sample click information to obtain the trained second factorization machine model; the trained first factorization machine model and the second factorization machine model constitute a recall system.

15. The apparatus according to claim 14, wherein the acquisition module is specifically used for: Obtain multiple sets of initial samples, wherein the initial samples include the user identifier, the content identifier, and the sample click information; Obtain at least one user profile of the sample user and at least one content tag of the sample content; The first sample is obtained based on the at least one user profile, the content identifier, and the sample click information; The second sample is obtained based on the user identifier, the at least one content tag, and the sample click information.

16. The apparatus according to claim 14 or 15, wherein the first sample comprises a first training sample; the first training module is specifically used for: Input the user profile and content identifier from the first training sample into the first factorization machine model to obtain the user profile vector and content identifier vector output by the first factorization machine model. The first click information is obtained based on the user profile vector and the content identifier vector.

17. The apparatus of claim 16, wherein the first sample further comprises a first verification sample; the first training module is further configured to: The user profile and content identifier in the first verification sample are input into the first factorization machine model to obtain the click information output by the first factorization machine model. Based on the click information output by the first factorization machine model and the corresponding sample click information, the area under the first ROC curve of the first factorization machine model is obtained. The first verification result is determined based on the area under the first ROC curve. The first verification result is used to indicate whether the trained first factorization machine model meets the requirements.

18. The apparatus of claim 17, wherein if the first verification result indicates that the trained first factorization machine model does not meet the requirements, the first training module is further configured to: Update the hyperparameters of the first factorization machine model, and retrain the first factorization machine model based on the updated hyperparameters and the first training samples.

19. The apparatus according to claim 14 or 15, wherein the second sample includes a second training sample; the second training module is specifically used for: Input the user identifier and content label from the second training sample into the second factorization machine model to obtain the user identifier vector and content label vector output by the second factorization machine model; The second click information is obtained based on the user identifier vector and the content tag vector.

20. The apparatus of claim 19, wherein the second sample further comprises a second verification sample; the second training module is further configured to: The user identifier and content tag in the second verification sample are input into the second factorization machine model to obtain the click information output by the second factorization machine model; Based on the click information output by the second factorization machine model and the corresponding sample click information, the area under the second ROC curve of the second factorization machine model is obtained; The second verification result is determined based on the area under the second ROC curve. The second verification result is used to indicate whether the trained second factorization machine model meets the requirements.

21. The apparatus of claim 20, wherein if the second verification result indicates that the trained second factorization machine model does not meet the requirements, the second training module is further configured to: Update the hyperparameters of the second factorization machine model, and retrain the second factorization machine model based on the updated hyperparameters and the second training samples.

22. A content recall device, comprising: The first processing module is used to obtain the target user profile vector of the target user and the target content identifier vector of multiple target contents according to the first factor decomposition machine model. The second processing module is used to obtain the target user identifier vector of the target user and the target content tag vector of the multiple target contents according to the second factorization machine model. The recall module is used to determine the recall content from the plurality of target contents based on the target user profile vector, the target content identifier vector, the target user identifier vector, and the target content tag vector; The first factorization machine model and the second factorization machine model are trained based on the model training method as described in any one of claims 1-8.

23. The apparatus according to claim 22, wherein the recall module is specifically used for: The target user profile vector and the target user identifier vector are concatenated to obtain the target user vector of the target user. For any content among the multiple target contents, the target content identifier vector and the target content tag vector of the content are concatenated to obtain the target content vector of the content. The recall content is determined from the plurality of target contents based on the target user vector and the target content vector of each of the contents.

24. The apparatus according to claim 23, wherein the recall module is specifically used for: Perform an inner product operation on the target user vector and the target content vector of each content to obtain the inner product score of each content for the target user; Based on the inner product score, a preset number of target contents are determined from the plurality of contents as the recall contents.

25. The apparatus according to any one of claims 22-24, wherein the first factorization machine model includes parameters for each user profile and parameters for each content identifier; the first processing module is specifically used for: Obtain the target user profile of the target user and the target content identifier of the target content; The target user profile vector is obtained based on the target user profile and the parameters corresponding to the target user profile in the first factorization machine model. The target content identifier vector is obtained based on the target content identifier and the parameters corresponding to the target content identifier in the first factorization machine model.

26. The apparatus according to any one of claims 22-24, wherein the second factorization machine model includes parameters for each user identifier and parameters for each content tag; the second processing module is specifically used for: Obtain the target user identifier of the target user and the target content tag of the target content; The target user identifier vector is obtained based on the target user identifier and the parameters corresponding to the target user identifier in the second factorization machine model; The target content tag vector is obtained based on the target content tag and the parameters corresponding to the target content tag in the second factorization machine model.

27. A computing device, comprising: At least one processor and memory; The memory stores computer-executed instructions; The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the model training method as described in any one of claims 1-8, or causing the at least one processor to perform the content retrieval method as described in any one of claims 9-13.

28. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the model training method as described in any one of claims 1-8, or implement the content retrieval method as described in any one of claims 9-13.

29. A computer program product comprising a computer program; wherein, when executed, the computer program implements the model training method according to any one of claims 1-8, or, when executed, the computer program implements the content recall method according to any one of claims 9-13.

Images